01 / On a schedule
Put a check on the clock
Nightly checks, weekly reconciliations, hourly monitors. Each run’s results land in its space, and the team is told.
/agents
Agents that do the recurring work.
Describe a job your team repeats. The agent names itself, suggests when to run, and sends you what it finds.
Approved knowledge. Read-only by default. Every step logged.
check 2026 RADV diagnoses, flag what lacks support, report by 6am
Findings · delivered 2:00 AM
18diagnoses without support, by provider
- 4 agents · 9 queries
- 1,240 checked
- 0 writes
01 · Create
Describe the job.
It sets itself up.
Click New bot and type the job in plain words. It names itself from your first message and suggests routines for the job. Add the ones you want.
Tell it what to do.
Every morning, check yesterday’s paid claims for duplicates. Show me the biggest first.
DW
I can run that. Each morning I’ll compare yesterday’s paid claims with claims for the same member, provider, date of service, and code.
I suggest two routines. Add the ones you want.
Message Duplicate WatchEvery morning, check yesterday’s paid claims for duplicates. Show me the biggest first.
01 · One agent, at work
Give it a task.
Follow the work.
See what it read, why it checked again, and what it found.
Reconcile March paid claims with the general ledger. Explain the difference.
Date cutoff mismatch
$12,400 falls outside the cutoff.
18 claims adjudicated March 31 were paid in the April 2 check run. The general ledger books them in April.
- Claims system, March
- $1,296,400
- By check date
- $1,284,000Matches the general ledger.
Read-only · source records unchanged
Run replay
Every step recorded.
02 · How it works
Connect. Work. Deliver.
Agents start on a schedule, not a question. They read through the connections you approved, work in steps, and deliver the work with a receipt for every query.
Every night · 2:00 AM
SQL Server · claims_dw
Encounters
Provider master
1,284 tables · mapped
Work · Run 2a173 of 3 agents done
Pick a lane to see its receipt
Delivered · Run 2a17Risk adjustment
18 diagnoses lack support
- 11No encounter on the date of service
- 5Provider type CMS does not accept
- 2Diagnosis missing from the encounter
- Ticket CDI-447118 charts to the coding queue
- ReportTo the risk adjustment lead
- Next runTomorrow at 2:00 AM
9 queries · 12s · 0 writes · all logged
Receipt · Run 2a17
- 02:00:00schedule · nightly RADV check · startedtrigger
- 02:00:01connect · claims_dw · SQL Serverok
- 02:00:01connect · encountersok
- 02:00:02connect · provider_masterok
- 02:00:04recalled · 4 approved pitfallsknowledge
- 02:00:07diagnoses agent · pull 1,240 submitted2.1s · logged
- 02:00:09encounters agent · match on date of service2.6s · logged
- 02:00:11providers agent · check accepted types1.8s · logged
- 02:00:12finding · 18 diagnoses without supportfound
- 02:00:12ticket CDI-4471 · 18 charts to the coding queueopened
- 02:00:12report · to the risk adjustment leadsent
- 02:00:12next run · tomorrow 02:00scheduled
- 02:00:12audit trail · 13 entries · 0 writescomplete
03 · Ways to run them
Put agents to work your way.
A few of the ways teams run them. Schedule the work, start from a question, or bring the agents you already use.
02 / From a question
Ask, and it does the work
An agent plans the question, runs it across your systems, and checks the answer. Save it, and it becomes a routine.
03 / Your own agents
Bring the AI you already run
Your assistants and internal tools read the same approved knowledge, so they answer from what your team agreed.
Try a task
Access needed
This source is outside the approved connections.
The agent stops here and says where the data sits. This source was not queried.
CMS CAP risk checkScope
04 · Guardrails
Read-only by default. Honest about its limits.
A new bot can read its connections and change nothing. It reaches only what it was granted. More access is requested with a reason and ends on its own.
Asked for something outside its scope, it says where the data sits and stops there.
05 · It learns
What one run learns, the next can use.
Agents propose findings from their work. Approved findings join the knowledge your team shares.
Your team reviews them, or delegates review to an agent under rules you set. Each approval records who made it and when.
Nightly reconciliation
Reading claims.claim_line
Proposed pitfall
Reversed lines count twice in paid totals.
claims.claim_line · 214 reversed lines
Later run · 2:00 AM
Paid total, March
Waiting for its next run
1,284approved notes
Approved onceUsed by every run
- Found
- Under review
- Approved
- Reused
06 · Questions
What your team will ask.
What can an agent do on its own?
Read the connections in its scope, run its routine, check its work, and notify your team. A new bot is read-only by default.
How is an agent kept in scope?
It works on approved, time-limited access to the connections you grant it. Asked for data outside that scope, it says where the data sits and what it would need, instead of guessing.
Is every step logged?
Yes. Every query an agent runs lands on the audit trail, and each run’s report carries its queries, rows, and checks, so your team can test it against the source.
Can our own AI tools use this?
Yes. The assistants and internal tools you already run, and any MCP client, can read the same approved knowledge your team reviews.
Do agents change the knowledge on their own?
They propose what they learn. Early on, a person approves each change; as proposals hold up, agents approve routine ones themselves. Every change goes into the knowledge graph with who approved it, and syncs to your Git repo for review.
Put your first check on a schedule.
Pick the check your team runs by hand every week. We connect the systems it needs and set it to run, read-only, with every step logged.